This thesis examines the relationship between large language models (LLMs) and political polarization through a reasoned review of recent literature. Starting from classical definitions of polarization, media bias, affective polarization, and framing, the work analyzes LLMs as relevant systems in the processes of production, selection, and mediation of information. The review organizes the selected contributions along three main levels: the outputs generated by the models, the effects on users, and the collective dynamics observable in multi-agent systems or simulated information environments. The analysis shows that polarization associated with LLMs cannot be reduced solely to the presence of bias in their outputs, but depends on the interaction between training data, alignment procedures, conversational form, source perception, and the structure of the information environment. Particular attention is devoted to the measures used in the literature, their limits in terms of reliability and validity, and mitigation strategies, including auditing, interaction design, content diversification, and transparency. In conclusion, LLMs appear to be ambivalent tools: they may contribute to amplifying polarizing dynamics, but under specific conditions they may also foster greater openness toward opposing perspectives. A psychological perspective is therefore essential to understand how these systems affect trust, persuasion, social identity, and the formation of public opinion.
L’elaborato esamina il rapporto tra large language models (LLM) e polarizzazione politica attraverso una rassegna ragionata della letteratura recente. A partire dalle definizioni classiche di polarizzazione, bias mediatico, polarizzazione affettiva e framing, il lavoro analizza gli LLM come sistemi rilevanti nei processi di produzione, selezione e mediazione dell’informazione. La rassegna organizza i contributi considerati lungo tre livelli principali: gli output generati dai modelli, gli effetti sugli utenti e le dinamiche collettive osservabili in sistemi multi-agente o ambienti informativi simulati. Dall’analisi emerge che la polarizzazione associata agli LLM non può essere ricondotta soltanto alla presenza di bias negli output, ma dipende dall’interazione tra dati di addestramento, procedure di allineamento, forma conversazionale, percezione della fonte e struttura dell’ambiente informativo. Particolare attenzione è dedicata alle misure utilizzate nella letteratura, ai loro limiti di affidabilità e validità, e alle strategie di mitigazione, tra cui auditing, progettazione dell’interazione, diversificazione dei contenuti e trasparenza. In conclusione, gli LLM appaiono come strumenti ambivalenti: possono contribuire ad amplificare dinamiche polarizzanti, ma in specifiche condizioni possono anche favorire una maggiore apertura verso prospettive opposte. Una prospettiva psicologica risulta quindi essenziale per comprendere come tali sistemi incidano su fiducia, persuasione, identità sociale e formazione dell’opinione pubblica.
Large language models e polarizzazione politica: una rassegna ragionata della letteratura
CORRADIN, GIOVANNI
2025/2026
Abstract
This thesis examines the relationship between large language models (LLMs) and political polarization through a reasoned review of recent literature. Starting from classical definitions of polarization, media bias, affective polarization, and framing, the work analyzes LLMs as relevant systems in the processes of production, selection, and mediation of information. The review organizes the selected contributions along three main levels: the outputs generated by the models, the effects on users, and the collective dynamics observable in multi-agent systems or simulated information environments. The analysis shows that polarization associated with LLMs cannot be reduced solely to the presence of bias in their outputs, but depends on the interaction between training data, alignment procedures, conversational form, source perception, and the structure of the information environment. Particular attention is devoted to the measures used in the literature, their limits in terms of reliability and validity, and mitigation strategies, including auditing, interaction design, content diversification, and transparency. In conclusion, LLMs appear to be ambivalent tools: they may contribute to amplifying polarizing dynamics, but under specific conditions they may also foster greater openness toward opposing perspectives. A psychological perspective is therefore essential to understand how these systems affect trust, persuasion, social identity, and the formation of public opinion.| File | Dimensione | Formato | |
|---|---|---|---|
|
Corradin_Giovanni_2013061.pdf
accesso aperto
Dimensione
578.43 kB
Formato
Adobe PDF
|
578.43 kB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/113865